I’m working on developing a differentiable CFD solver incorporating the immersed boundary method for solving the Navier-Stokes equations. I plan on using Enzyme.jl .My goal is to compute gradients of a parametric geometry with respect to a custom-defined loss function, which I plan to optimize using gradient descent. However, I’m encountering memory issues due to the large computational graph resulting from multiple iterations (i.e., time steps). Are there any effective strategies to mitigate this problem?
For time stepping adjoints, you normally want to solve this by not differentiating the solver directly but instead define some adjoints over it that optimize a few things. Some of things you can optimize are memory with continuous checkpointing and compute time because some of the computations can be skipped. If you use DifferentialEquations.jl for the time stepping or NonlinearSolve.jl for the nonlinear solving these optimizations will be automatically applied.
